fix: add serveBaseUrl to compactEmbeddedPiSession params

2026-01-07 06:00:21 +00:00
parent 1a47aec6e4
commit f85807a2a6
35 changed files with 3427 additions and 1 deletions
--- a/skills/thebuilders-v2/LESSONS.md
+++ b/skills/thebuilders-v2/LESSONS.md
@@ -0,0 +1,197 @@
+# Lessons Learned - The Builders v2
+
+Post-mortem from the OpenAI episode generation session (Jan 6, 2026).
+
+## Issues & Fixes
+
+### 1. Wrong Model Used
+**Issue:** Used GPT-4o instead of GPT-5.2 as requested  
+**Result:** 47 turns instead of 120+  
+**Fix:** Always confirm model before running. Add `-m` flag to CLI.
+
+### 2. Script Too Short in Single Pass
+**Issue:** Even GPT-5.2 only generated 90 turns when asked for 120+  
+**Result:** Episode too shallow (7 min instead of 25 min)  
+**Fix:** **Section-by-section generation**. LLMs follow length instructions better on smaller chunks with explicit turn targets.
+
+### 3. TTS Chunked by Size, Not by Speaker
+**Issue:** Audio chunks alternated voices randomly instead of per speaker  
+**Result:** Sounded like a confused monologue  
+**Fix:** Parse script by `<Person1>`/`<Person2>` tags, assign consistent voice per speaker.
+
+### 4. v3 Style Guide Not Integrated
+**Issue:** The Acquired Bible existed but wasn't in the prompts  
+**Result:** Generic dialogue, not Acquired-style discovery  
+**Fix:** Include the Bible in EVERY section prompt. The patterns matter:
+- Interruptions ("Wait—", "Hold on—")
+- Discovery moments ("That's insane", "Wait, really?")
+- Distinct perspectives (Maya=technical, James=strategy)
+
+### 5. No Intro/Greeting
+**Issue:** Jumped straight into hook with no "Welcome to..."  
+**Result:** Felt robotic, not like real hosts  
+**Fix:** Add INTRO section (8 turns) with:
+- Welcome
+- Casual banter
+- "Today we're covering X"
+- Tease the story
+
+### 6. Voices Too Similar
+**Issue:** Used nova/onyx which sound somewhat similar  
+**Result:** Hard to distinguish hosts  
+**Fix:** Use more distinct voices: **alloy** (female) + **echo** (male)
+
+## What Works
+
+### Section-by-Section
+Each section has explicit turn targets:
+```
+INTRO:        8 turns
+HOOK:        12 turns
+ORIGIN:      20 turns
+INFLECTION1: 18 turns
+INFLECTION2: 18 turns
+MESSY_MIDDLE: 14 turns
+NOW:         12 turns
+TAKEAWAYS:   10 turns
+─────────────────────
+Total:      112 turns
+```
+
+### Speaker-Aware TTS
+```javascript
+// WRONG - chunks by size
+for (chunk of script.split(3500)) {
+  voice = i % 2 === 0 ? 'nova' : 'onyx';
+}
+
+// RIGHT - parses by speaker
+for (turn of script.matchAll(/<(Person[12])>(.+?)<\/Person[12]>/)) {
+  voice = turn[1] === 'Person1' ? 'alloy' : 'echo';
+}
+```
+
+### Bible in Every Prompt
+The `acquired-bible.md` template contains:
+- Host personalities
+- Conversation patterns
+- Language to use
+- What NOT to do
+
+Including it in every section prompt ensures consistency.
+
+## Pipeline Summary
+
+```
+1. Gemini Deep Research (~5 min)
+2. Hook Generation (~15s)
+3. Section Generation (7 sections × ~20s = ~2.5 min)
+4. Speaker-Aware TTS (~45s for 112 turns)
+5. FFmpeg Merge (~2s)
+────────────────────────────────────
+Total: ~8-10 min for 20-25 min episode
+```
+
+### 7. Facts Repeated Across Sections
+**Issue:** Same facts (hot dog $1.50, renewal rate 93.3%, etc.) repeated 10-20 times  
+**Costco example:** Hot dog mentioned 19×, renewal rate 20×, chicken 15×  
+**Root cause:** Each section generated independently with same research context  
+**Fix:** **Deduplication system**
+- Extract key facts after each section
+- Pass "DO NOT REPEAT" list to subsequent sections
+- Track up to 100 facts across sections
+
+### 8. Process Dies Mid-Generation
+**Issue:** Long-running generation killed by OOM or gateway timeout  
+**Result:** Lost 30+ minutes of work, had to restart  
+**Fix:** **Checkpoint system**
+- Save each section immediately after generation
+- Save state (usedFacts, prevContext) to JSON
+- On restart, detect checkpoint and resume from last section
+- Location: `<output>/checkpoints/`
+
+## What Works
+
+### Checkpoint System
+```
+<output>/checkpoints/
+├── section-0.txt   # INTRO
+├── section-1.txt   # HOOK
+├── ...
+└── state.json      # { completedSections: 5, usedFacts: [...] }
+```
+
+If process dies at section 7, re-run same command → resumes from section 7.
+
+### Deduplication
+```
+Section 0: Extract 17 facts → pass to Section 1 as "DO NOT USE"
+Section 1: Extract 16 facts → 33 total blocked
+Section 2: Extract 14 facts → 47 total blocked
+...
+```
+
+Result: 100 unique facts tracked, no repetition.
+
+### Section-by-Section
+Each section has explicit turn targets:
+```
+INTRO:        8 turns
+HOOK:        12 turns
+ORIGIN:      20 turns
+INFLECTION1: 18 turns
+INFLECTION2: 18 turns
+MESSY_MIDDLE: 14 turns
+NOW:         12 turns
+TAKEAWAYS:   10 turns
+─────────────────────
+Total:      112 turns
+```
+
+### Speaker-Aware TTS
+```javascript
+// WRONG - chunks by size
+for (chunk of script.split(3500)) {
+  voice = i % 2 === 0 ? 'nova' : 'onyx';
+}
+
+// RIGHT - parses by speaker
+for (turn of script.matchAll(/<(Person[12])>(.+?)<\/Person[12]>/)) {
+  voice = turn[1] === 'Person1' ? 'alloy' : 'echo';
+}
+```
+
+### Bible in Every Prompt
+The `acquired-bible.md` template contains:
+- Host personalities
+- Conversation patterns
+- Language to use
+- What NOT to do
+
+Including it in every section prompt ensures consistency.
+
+## Pipeline Summary
+
+```
+1. Gemini Deep Research (~5 min)
+2. Hook Generation (~15s)
+3. Section Generation (11 sections × ~25s = ~4.5 min) [with checkpoints]
+4. Speaker-Aware TTS (~2 min for 250+ turns)
+5. FFmpeg Merge (~2s)
+────────────────────────────────────
+Total: ~12-15 min for 45-60 min deep dive
+       ~8-10 min for 20 min quick dive
+```
+
+## Checklist for Future Episodes
+
+- [ ] Confirm model (gpt-5.2 or better)
+- [ ] Run deep research first
+- [ ] Generate section-by-section
+- [ ] Include Bible in all prompts
+- [ ] Parse by speaker tags for TTS
+- [ ] Use alloy + echo voices
+- [ ] Verify intro section exists
+- [ ] Listen to first 30s to check voice distinction
+- [ ] Verify no repeated facts (deduplication working)
+- [ ] If process dies, just re-run same command (checkpoint resume)
--- a/skills/thebuilders-v2/SKILL.md
+++ b/skills/thebuilders-v2/SKILL.md
@@ -0,0 +1,111 @@
+# The Builders Podcast Generator v2
+
+Generate Acquired-style podcast episodes with two distinct hosts.
+
+## Two Modes
+
+| Mode | Flag | Duration | Research | Turns | Use Case |
+|------|------|----------|----------|-------|----------|
+| **Quick Dive** | `--quick` | 15-20 min | 1 pass Gemini | ~110 | Fast turnaround |
+| **Deep Dive** | `--deep` | 45-60 min | 5-layer + o3 synthesis | ~250 | True Acquired-style |
+
+## Quick Start
+
+```bash
+cd /home/elie/github/clawdis/skills/thebuilders-v2
+
+# Quick dive (15-20 min) - default
+node scripts/generate.mjs "Costco" -o /tmp/costco
+
+# Deep dive (45-60 min)
+node scripts/generate.mjs "Costco" --deep -o /tmp/costco-deep
+
+# Skip research (use existing)
+node scripts/generate.mjs "Costco" --deep -o /tmp/costco --skip-research
+```
+
+## Quick Dive Mode
+
+Single research pass, 8 sections:
+- INTRO (8 turns)
+- HOOK (12 turns)
+- ORIGIN (20 turns)
+- INFLECTION_1 (18 turns)
+- INFLECTION_2 (18 turns)
+- MESSY_MIDDLE (14 turns)
+- NOW (12 turns)
+- TAKEAWAYS (10 turns)
+
+**Total: ~112 turns / 15-20 min**
+
+## Deep Dive Mode
+
+5-layer research + 11 sections:
+
+### Research Layers
+1. **Overview** - Gemini Deep Research
+2. **Founders** - Interview quotes, early stories
+3. **Decisions** - Contemporary WSJ/NYT coverage
+4. **Financials** - Revenue, margins, metrics
+5. **Contrarian** - Failures, criticism, struggles
+
+### Research Synthesis (o3)
+Extracts:
+- 50 most surprising facts
+- 10 best quotes with attribution
+- 5 "Wait, really?" moments
+- Key timeline with dates
+- Contrarian takes
+
+### Sections
+- INTRO (10 turns)
+- HOOK (15 turns)
+- ORIGIN (35 turns) - deep founder story
+- EARLY_PRODUCT (25 turns) - how it actually worked
+- INFLECTION_1 (25 turns) - first major decision
+- SCALE (25 turns) - operational details
+- INFLECTION_2 (25 turns) - second crisis
+- COMPETITION (20 turns)
+- CULTURE (20 turns)
+- NOW (20 turns)
+- TAKEAWAYS (15 turns)
+
+**Total: ~235 turns / 45-60 min**
+
+## The Hosts
+
+| Host | Voice | Perspective |
+|------|-------|-------------|
+| **Maya Chen** (Person1) | alloy | Technical - "How does this actually work?" |
+| **James Porter** (Person2) | echo | Strategic - "For context... The lesson here is..." |
+
+## Output Files
+
+```
+output/
+├── research.md              # Quick mode research
+├── research-combined.md     # Deep mode combined research
+├── research-synthesis.md    # Deep mode o3 synthesis
+├── hooks.md                 # Hook options
+├── script.txt               # Full dialogue
+├── audio/                   # Individual turns
+└── episode.mp3              # Final episode
+```
+
+## Time Estimates
+
+| Mode | Research | Script | TTS | Total |
+|------|----------|--------|-----|-------|
+| Quick | 5 min | 3 min | 1 min | ~10 min |
+| Deep | 15 min | 10 min | 3 min | ~30 min |
+
+## Key Differences
+
+| Aspect | Quick | Deep |
+|--------|-------|------|
+| Research passes | 1 | 5 |
+| Research synthesis | None | o3 extracts key facts |
+| Sections | 8 | 11 |
+| ORIGIN depth | 20 turns | 35 turns |
+| Quotes required | Encouraged | Mandatory |
+| Context per section | 10KB | 25KB |
--- a/skills/thebuilders-v2/config/presets.yaml
+++ b/skills/thebuilders-v2/config/presets.yaml
@@ -0,0 +1,59 @@
+# Episode Presets
+
+company_deep_dive:
+  description: "Full company history deep dive"
+  target_length: 25  # minutes
+  sections:
+    - hook
+    - origin
+    - inflection_1
+    - inflection_2
+    - messy_middle
+    - now
+    - takeaways
+  hook_style: story
+  citation_density: medium
+  min_turns: 120
+
+quick_explainer:
+  description: "Quick overview of a topic"
+  target_length: 10
+  sections:
+    - hook
+    - what
+    - why
+    - how
+    - takeaway
+  hook_style: question
+  citation_density: low
+  min_turns: 50
+
+founder_story:
+  description: "Focus on founder journey"
+  target_length: 20
+  sections:
+    - hook
+    - before
+    - genesis
+    - struggle
+    - breakthrough
+    - now
+    - lessons
+  hook_style: story
+  citation_density: medium
+  min_turns: 100
+
+product_breakdown:
+  description: "Technical product deep dive"
+  target_length: 20
+  sections:
+    - hook
+    - problem
+    - solution
+    - architecture
+    - evolution
+    - impact
+    - future
+  hook_style: data
+  citation_density: high
+  min_turns: 100
--- a/skills/thebuilders-v2/config/voices.yaml
+++ b/skills/thebuilders-v2/config/voices.yaml
@@ -0,0 +1,47 @@
+# Host Voice Configuration
+
+maya:
+  tag: Person1
+  role: Technical Host
+  tts_voice: nova  # OpenAI voice
+  
+  personality:
+    background: "Former software engineer, B2B SaaS founder"
+    lens: "Technical, mechanical, architectural"
+    style: "Direct, skeptical of hype, loves clever solutions"
+  
+  signature_phrases:
+    - "Wait, slow down."
+    - "Show me the numbers."
+    - "I had to read this three times."
+    - "That's actually clever because..."
+    - "How does that actually work?"
+    - "Let's look at the architecture."
+  
+  never_says:
+    - "From a strategic perspective"
+    - "The market opportunity"
+    - "In terms of positioning"
+
+james:
+  tag: Person2
+  role: Strategy Host
+  tts_voice: onyx  # OpenAI voice
+  
+  personality:
+    background: "Former VC, business history student"
+    lens: "Strategy, markets, competitive dynamics"
+    style: "Synthesizer, pattern matcher, historical parallels"
+  
+  signature_phrases:
+    - "For context..."
+    - "This is the classic [X] playbook."
+    - "The lesson here is..."
+    - "What a story."
+    - "And this is where things get messy."
+    - "The endgame here is..."
+  
+  never_says:
+    - "Let me explain the architecture"
+    - "The API design"
+    - "From a technical standpoint"
--- a/skills/thebuilders-v2/scripts/generate-sections.mjs
+++ b/skills/thebuilders-v2/scripts/generate-sections.mjs
@@ -0,0 +1,271 @@
+#!/usr/bin/env node
+/**
+ * Section-by-section podcast script generator
+ * Generates each section separately to ensure proper length
+ */
+
+import { readFileSync, writeFileSync, mkdirSync } from 'fs';
+import { join, dirname } from 'path';
+import { fileURLToPath } from 'url';
+import { execSync } from 'child_process';
+
+const __dirname = dirname(fileURLToPath(import.meta.url));
+const OPENAI_API_KEY = process.env.OPENAI_API_KEY;
+
+// Section definitions with turn targets
+const SECTIONS = [
+  { id: 1, name: 'HOOK', turns: 15, minutes: '3-4', description: 'Opening hook, establish central question, create curiosity' },
+  { id: 2, name: 'ORIGIN', turns: 25, minutes: '6-7', description: 'Founders, genesis, early bet, market context, human element' },
+  { id: 3, name: 'INFLECTION_1', turns: 20, minutes: '5-6', description: 'First major decision point, alternatives, stakes, outcome' },
+  { id: 4, name: 'INFLECTION_2', turns: 20, minutes: '5-6', description: 'Second pivot, new challenge, adaptation, insight' },
+  { id: 5, name: 'MESSY_MIDDLE', turns: 15, minutes: '3-4', description: 'Near-death moments, internal conflicts, survival' },
+  { id: 6, name: 'NOW', turns: 15, minutes: '3-4', description: 'Current position, competition, open questions' },
+  { id: 7, name: 'TAKEAWAYS', turns: 12, minutes: '2-3', description: 'Key lessons, frameworks, final thought' },
+];
+
+const TOTAL_TURNS = SECTIONS.reduce((sum, s) => sum + s.turns, 0);
+
+const log = (msg) => console.log(`[${new Date().toISOString().slice(11,19)}] ${msg}`);
+
+async function llm(prompt, model = 'gpt-5.2') {
+  const response = await fetch('https://api.openai.com/v1/chat/completions', {
+    method: 'POST',
+    headers: {
+      'Authorization': `Bearer ${OPENAI_API_KEY}`,
+      'Content-Type': 'application/json',
+    },
+    body: JSON.stringify({
+      model,
+      messages: [{ role: 'user', content: prompt }],
+      max_completion_tokens: 8000,
+      temperature: 0.7,
+    }),
+  });
+
+  if (!response.ok) {
+    const error = await response.text();
+    throw new Error(`API error: ${response.status} - ${error}`);
+  }
+
+  const data = await response.json();
+  return data.choices[0].message.content;
+}
+
+function buildSectionPrompt(section, research, hook, previousSections) {
+  const hostsInfo = `
+## The Hosts
+
+**Maya Chen (Person1):**
+- Former software engineer, B2B SaaS founder
+- Lens: "How does this actually work?" - technical, mechanical
+- Skeptical of hype, wants specifics and numbers
+- Phrases: "Wait, slow down.", "Show me the numbers.", "That's actually clever because..."
+
+**James Porter (Person2):**
+- Former VC, business history student  
+- Lens: "What's the business model?" - strategy, markets, competitive dynamics
+- Synthesizer, pattern matcher, historical parallels
+- Phrases: "For context...", "This is the classic [X] playbook.", "The lesson here is..."
+`;
+
+  const formatRules = `
+## Format Rules
+- Use <Person1> and <Person2> XML tags for each speaker
+- Each turn: 2-5 sentences (natural conversation, not monologues)
+- Include discovery moments: "Wait, really?", "I had no idea!"
+- Use SPECIFIC numbers, dates, dollar amounts from research
+- Both hosts should react naturally and build on each other
+`;
+
+  const previousContext = previousSections.length > 0 
+    ? `\n## Previous Sections (for context, don't repeat):\n${previousSections.join('\n\n')}\n`
+    : '';
+
+  const sectionSpecific = {
+    HOOK: `Open with this hook and build on it:\n"${hook}"\n\nEstablish why this story matters. Create curiosity about what's coming.`,
+    ORIGIN: `Cover the founding story. Who are the founders? What sparked this? What was their early bet? Make them human and relatable.`,
+    INFLECTION_1: `Cover the FIRST major decision/pivot point. What choice did they face? What alternatives existed? What did they risk? How did it play out?`,
+    INFLECTION_2: `Cover the SECOND major inflection point. What new challenge emerged? How did they adapt? What insight did they gain?`,
+    MESSY_MIDDLE: `Cover the struggles. What almost killed them? What internal conflicts existed? Don't glorify - show real struggle.`,
+    NOW: `Cover current state. Where are they now? Key metrics? Competitive position? What questions remain open?`,
+    TAKEAWAYS: `Wrap up with lessons learned. What's the key framework? What would you do differently? End with a memorable final thought that ties back to the hook.`,
+  };
+
+  return `# Generate Section ${section.id}: ${section.name}
+
+${hostsInfo}
+
+## Your Task
+Generate EXACTLY ${section.turns} dialogue turns for the ${section.name} section.
+Target duration: ${section.minutes} minutes when read aloud.
+
+## Section Goal
+${section.description}
+
+## Specific Instructions
+${sectionSpecific[section.name]}
+
+${formatRules}
+
+${previousContext}
+
+## Research Material
+${research}
+
+---
+
+Generate EXACTLY ${section.turns} dialogue turns. Start with <!-- SECTION ${section.id}: ${section.name} --> then <Person1> or <Person2>.
+Count your turns carefully - you MUST hit ${section.turns} turns.`;
+}
+
+async function generateSection(section, research, hook, previousSections) {
+  log(`Generating Section ${section.id}: ${section.name} (target: ${section.turns} turns)...`);
+  
+  const prompt = buildSectionPrompt(section, research, hook, previousSections);
+  const startTime = Date.now();
+  
+  const result = await llm(prompt);
+  
+  const elapsed = ((Date.now() - startTime) / 1000).toFixed(1);
+  const turns = (result.match(/<Person[12]>/g) || []).length;
+  
+  log(`Section ${section.id} complete: ${turns} turns in ${elapsed}s`);
+  
+  return { section, result, turns };
+}
+
+async function main() {
+  const args = process.argv.slice(2);
+  const outputDir = args.includes('-o') ? args[args.indexOf('-o') + 1] : './output';
+  
+  console.log(`
+╔════════════════════════════════════════════════════════════╗
+║  Section-by-Section Podcast Generator                      ║
+║  Target: ${TOTAL_TURNS} dialogue turns                                  ║
+╚════════════════════════════════════════════════════════════╝
+`);
+
+  // Load research and hook
+  const research = readFileSync(join(outputDir, '02-research.md'), 'utf-8');
+  const hooksFile = readFileSync(join(outputDir, '03-hooks.md'), 'utf-8');
+  const hookMatch = hooksFile.match(/## Story Hook\n\n> "([^"]+)"/);
+  const hook = hookMatch ? hookMatch[1] : '';
+  
+  log(`Loaded research (${(research.length/1024).toFixed(1)}KB) and hook`);
+  log(`Hook: "${hook.slice(0, 60)}..."`);
+  
+  // Generate sections sequentially
+  const sections = [];
+  const previousSections = [];
+  let totalTurns = 0;
+  
+  for (const section of SECTIONS) {
+    const { result, turns } = await generateSection(section, research, hook, previousSections);
+    sections.push(result);
+    previousSections.push(result.slice(0, 500) + '...'); // Keep context manageable
+    totalTurns += turns;
+  }
+  
+  // Combine all sections
+  const fullScript = sections.join('\n\n');
+  writeFileSync(join(outputDir, '06-script-final.txt'), fullScript);
+  
+  log(`\n✅ Script complete: ${totalTurns} total turns`);
+  log(`Saved to ${join(outputDir, '06-script-final.txt')}`);
+  
+  // Generate audio
+  log('\nGenerating audio...');
+  
+  // Chunk the script
+  const chunksDir = join(outputDir, 'chunks');
+  mkdirSync(chunksDir, { recursive: true });
+  
+  const chunks = [];
+  let currentChunk = '';
+  for (const line of fullScript.split('\n')) {
+    if (currentChunk.length + line.length > 3500 && currentChunk.length > 1000) {
+      if (line.startsWith('<Person') || line.startsWith('<!--')) {
+        chunks.push(currentChunk.trim());
+        currentChunk = line + '\n';
+      } else {
+        currentChunk += line + '\n';
+      }
+    } else {
+      currentChunk += line + '\n';
+    }
+  }
+  if (currentChunk.trim()) chunks.push(currentChunk.trim());
+  
+  log(`Split into ${chunks.length} audio chunks`);
+  
+  // Save chunks
+  chunks.forEach((chunk, i) => {
+    const num = String(i + 1).padStart(3, '0');
+    writeFileSync(join(chunksDir, `chunk-${num}.txt`), chunk);
+  });
+  
+  // Generate TTS in parallel
+  const ttsStart = Date.now();
+  const audioPromises = chunks.map(async (chunk, i) => {
+    const num = String(i + 1).padStart(3, '0');
+    const text = chunk
+      .replace(/<Person1>/g, '')
+      .replace(/<Person2>/g, '')
+      .replace(/<!--[^>]+-->/g, '')
+      .trim()
+      .slice(0, 4096);
+    
+    const outPath = join(chunksDir, `chunk-${num}.mp3`);
+    
+    const res = await fetch('https://api.openai.com/v1/audio/speech', {
+      method: 'POST',
+      headers: {
+        'Authorization': `Bearer ${OPENAI_API_KEY}`,
+        'Content-Type': 'application/json',
+      },
+      body: JSON.stringify({
+        model: 'tts-1',
+        input: text,
+        voice: i % 2 === 0 ? 'nova' : 'onyx',
+      }),
+    });
+    
+    if (!res.ok) throw new Error(`TTS failed for chunk ${num}`);
+    
+    const buffer = await res.arrayBuffer();
+    writeFileSync(outPath, Buffer.from(buffer));
+    log(`Chunk ${num} audio done`);
+    return outPath;
+  });
+  
+  const audioPaths = await Promise.all(audioPromises);
+  log(`TTS complete in ${((Date.now() - ttsStart) / 1000).toFixed(1)}s`);
+  
+  // Merge with ffmpeg
+  const listFile = join(chunksDir, 'files.txt');
+  writeFileSync(listFile, audioPaths.map(p => `file '${p}'`).join('\n'));
+  
+  const episodePath = join(outputDir, 'episode.mp3');
+  execSync(`ffmpeg -y -f concat -safe 0 -i "${listFile}" -c copy "${episodePath}" 2>/dev/null`);
+  
+  // Get duration
+  const duration = execSync(`ffprobe -v error -show_entries format=duration -of default=noprint_wrappers=1:nokey=1 "${episodePath}"`, { encoding: 'utf-8' }).trim();
+  const minutes = (parseFloat(duration) / 60).toFixed(1);
+  
+  console.log(`
+╔════════════════════════════════════════════════════════════╗
+║  Generation Complete                                       ║
+╠════════════════════════════════════════════════════════════╣
+║  Total turns: ${String(totalTurns).padEnd(44)}║
+║  Duration: ${String(minutes + ' minutes').padEnd(47)}║
+║  Chunks: ${String(chunks.length).padEnd(49)}║
+╚════════════════════════════════════════════════════════════╝
+
+Output: ${episodePath}
+`);
+}
+
+main().catch(err => {
+  console.error('Error:', err.message);
+  process.exit(1);
+});
--- a/skills/thebuilders-v2/scripts/generate.mjs
+++ b/skills/thebuilders-v2/scripts/generate.mjs
@@ -0,0 +1,639 @@
+#!/usr/bin/env node
+/**
+ * The Builders Podcast Generator v2
+ * 
+ * Two modes:
+ * --quick : 15-20 min episode, single research pass
+ * --deep  : 45-60 min episode, multi-layer research + enhancement
+ */
+
+import { readFileSync, writeFileSync, mkdirSync, readdirSync, existsSync, rmSync } from 'fs';
+import { join, dirname } from 'path';
+import { fileURLToPath } from 'url';
+import { execSync } from 'child_process';
+
+const __dirname = dirname(fileURLToPath(import.meta.url));
+const SKILL_DIR = join(__dirname, '..');
+const OPENAI_API_KEY = process.env.OPENAI_API_KEY;
+
+// Load the Acquired Bible
+const BIBLE = readFileSync(join(SKILL_DIR, 'templates/acquired-bible.md'), 'utf-8');
+
+// Section definitions - QUICK mode
+const SECTIONS_QUICK = [
+  { id: 0, name: 'INTRO', turns: 8, description: 'Welcome, casual banter, introduce the company, tease why this story is fascinating' },
+  { id: 1, name: 'HOOK', turns: 12, description: 'Open with surprising fact or tension. Central question. Stakes.' },
+  { id: 2, name: 'ORIGIN', turns: 20, description: 'Founders as humans. Genesis moment. Early bet. Market context.' },
+  { id: 3, name: 'INFLECTION_1', turns: 18, description: 'First major decision. Alternatives. Stakes. What they risked.' },
+  { id: 4, name: 'INFLECTION_2', turns: 18, description: 'Second pivot. New challenge. How they adapted.' },
+  { id: 5, name: 'MESSY_MIDDLE', turns: 14, description: 'Near-death. Internal conflicts. Real struggle, not glorified.' },
+  { id: 6, name: 'NOW', turns: 12, description: 'Current state. Metrics. Competition. Open questions.' },
+  { id: 7, name: 'TAKEAWAYS', turns: 10, description: 'Key lessons. Frameworks. Final thought tying back to hook.' },
+];
+
+// Section definitions - DEEP mode (more turns, more sections)
+const SECTIONS_DEEP = [
+  { id: 0, name: 'INTRO', turns: 10, description: 'Welcome, what excited them about this research, why this company matters now' },
+  { id: 1, name: 'HOOK', turns: 15, description: 'Vivid opening scene. The moment that defines the company. Stakes made visceral.' },
+  { id: 2, name: 'ORIGIN', turns: 35, description: 'Deep founder story - actual quotes, specific moments, human struggles. Market context of the era. What the world looked like.' },
+  { id: 3, name: 'EARLY_PRODUCT', turns: 25, description: 'First product/service. How it actually worked mechanically. Early customers. The insight that made it work.' },
+  { id: 4, name: 'INFLECTION_1', turns: 25, description: 'First major decision. Board meeting details. What alternatives existed. Quotes from people who were there.' },
+  { id: 5, name: 'SCALE', turns: 25, description: 'How they scaled. Operational details. What broke. How they fixed it. Specific numbers.' },
+  { id: 6, name: 'INFLECTION_2', turns: 25, description: 'Second pivot or crisis. The moment it almost fell apart. How they survived.' },
+  { id: 7, name: 'COMPETITION', turns: 20, description: 'Competitive landscape. Who they beat and how. Who almost beat them.' },
+  { id: 8, name: 'CULTURE', turns: 20, description: 'Internal culture. Leadership philosophy. Quotes from employees. What makes them different.' },
+  { id: 9, name: 'NOW', turns: 20, description: 'Current state with specific metrics. Recent moves. Open questions. What could go wrong.' },
+  { id: 10, name: 'TAKEAWAYS', turns: 15, description: 'Key frameworks. What other companies can learn. Final thought tying back to hook.' },
+];
+
+const VOICES = {
+  Person1: 'alloy',   // Maya
+  Person2: 'echo',    // James
+};
+
+const log = (msg) => console.log(`[${new Date().toISOString().slice(11,19)}] ${msg}`);
+
+// ============ LLM CALLS ============
+
+async function llm(prompt, model = 'gpt-5.2', maxTokens = 8000) {
+  const response = await fetch('https://api.openai.com/v1/chat/completions', {
+    method: 'POST',
+    headers: {
+      'Authorization': `Bearer ${OPENAI_API_KEY}`,
+      'Content-Type': 'application/json',
+    },
+    body: JSON.stringify({
+      model,
+      messages: [{ role: 'user', content: prompt }],
+      max_completion_tokens: maxTokens,
+      temperature: 0.8,
+    }),
+  });
+
+  if (!response.ok) {
+    const error = await response.text();
+    throw new Error(`API error: ${response.status} - ${error}`);
+  }
+
+  const data = await response.json();
+  return data.choices[0].message.content;
+}
+
+async function llmO3(prompt) {
+  // Use o3 for reasoning-heavy tasks
+  const response = await fetch('https://api.openai.com/v1/chat/completions', {
+    method: 'POST',
+    headers: {
+      'Authorization': `Bearer ${OPENAI_API_KEY}`,
+      'Content-Type': 'application/json',
+    },
+    body: JSON.stringify({
+      model: 'o3',
+      messages: [{ role: 'user', content: prompt }],
+      max_completion_tokens: 16000,
+    }),
+  });
+
+  if (!response.ok) {
+    const error = await response.text();
+    throw new Error(`o3 API error: ${response.status} - ${error}`);
+  }
+
+  const data = await response.json();
+  return data.choices[0].message.content;
+}
+
+// ============ RESEARCH ============
+
+async function runGeminiResearch(topic) {
+  log('Running Gemini Deep Research...');
+  const start = Date.now();
+  
+  const result = execSync(
+    `cd /home/elie/github/clawdis/skills/deepresearch && ./scripts/deepresearch.sh "${topic}"`,
+    { encoding: 'utf-8', maxBuffer: 50 * 1024 * 1024, timeout: 600000 }
+  );
+  
+  const elapsed = ((Date.now() - start) / 1000).toFixed(1);
+  log(`Gemini research complete in ${elapsed}s`);
+  
+  return result;
+}
+
+async function runBraveSearch(query, count = 5) {
+  try {
+    const result = execSync(
+      `node /home/elie/github/clawdis/skills/brave-search/scripts/search.mjs "${query}" -n ${count} --content 2>/dev/null`,
+      { encoding: 'utf-8', maxBuffer: 10 * 1024 * 1024, timeout: 60000 }
+    );
+    return result;
+  } catch (e) {
+    return '';
+  }
+}
+
+async function runExaSearch(query) {
+  try {
+    const result = execSync(
+      `node /home/elie/github/clawdis/skills/researcher/scripts/research.mjs "${query}" 2>/dev/null`,
+      { encoding: 'utf-8', maxBuffer: 10 * 1024 * 1024, timeout: 120000 }
+    );
+    return result;
+  } catch (e) {
+    return '';
+  }
+}
+
+async function multiLayerResearch(topic, company, outputDir) {
+  log('Starting 5-layer deep research...');
+  const research = {};
+  
+  // Layer 1: Overview (Gemini)
+  log('  Layer 1/5: Overview (Gemini)...');
+  research.overview = await runGeminiResearch(topic);
+  writeFileSync(join(outputDir, 'research-1-overview.md'), research.overview);
+  
+  // Layer 2: Founder stories
+  log('  Layer 2/5: Founder interviews...');
+  research.founders = await runBraveSearch(`"${company}" founder interview quotes early days story`, 8);
+  writeFileSync(join(outputDir, 'research-2-founders.md'), research.founders);
+  
+  // Layer 3: Key decisions (contemporary coverage)
+  log('  Layer 3/5: Contemporary coverage...');
+  research.decisions = await runBraveSearch(`"${company}" site:wsj.com OR site:nytimes.com OR site:forbes.com history`, 8);
+  writeFileSync(join(outputDir, 'research-3-decisions.md'), research.decisions);
+  
+  // Layer 4: Financial/operational details
+  log('  Layer 4/5: Financial details...');
+  research.financials = await runBraveSearch(`"${company}" revenue profit margin business model analysis`, 5);
+  writeFileSync(join(outputDir, 'research-4-financials.md'), research.financials);
+  
+  // Layer 5: Contrarian/struggles
+  log('  Layer 5/5: Struggles and criticism...');
+  research.contrarian = await runBraveSearch(`"${company}" almost failed crisis problems criticism`, 5);
+  writeFileSync(join(outputDir, 'research-5-contrarian.md'), research.contrarian);
+  
+  // Combine all research
+  const combined = `# RESEARCH COMPILATION: ${company}
+
+## PART 1: OVERVIEW
+${research.overview}
+
+## PART 2: FOUNDER STORIES & INTERVIEWS
+${research.founders}
+
+## PART 3: CONTEMPORARY COVERAGE & KEY DECISIONS
+${research.decisions}
+
+## PART 4: FINANCIAL & OPERATIONAL DETAILS
+${research.financials}
+
+## PART 5: STRUGGLES, CRISES & CRITICISM
+${research.contrarian}
+`;
+  
+  writeFileSync(join(outputDir, 'research-combined.md'), combined);
+  log('All research layers complete');
+  
+  return combined;
+}
+
+async function synthesizeResearch(research, company, outputDir) {
+  log('Synthesizing research with o3...');
+  
+  const prompt = `You are a research analyst preparing materials for a deep-dive podcast about ${company}.
+
+From this research, extract:
+
+1. **50 MOST SURPRISING FACTS** - specific numbers, dates, details that would make someone say "Wait, really?"
+
+2. **10 BEST QUOTES** - actual quotes from founders, employees, or articles WITH attribution
+   Format: "Quote here" — Person Name, Source, Year
+
+3. **5 "WAIT REALLY?" MOMENTS** - the most counterintuitive or shocking facts
+
+4. **KEY TIMELINE** - 15-20 most important dates with specific events
+
+5. **NARRATIVE THREADS** - the 3-4 main story arcs that make this company interesting
+
+6. **CONTRARIAN TAKES** - what critics say, what almost went wrong, the messy parts
+
+7. **NUMBERS THAT MATTER** - specific metrics that tell the story (revenue, margins, users, etc.)
+
+Be SPECIFIC. Include actual numbers, names, dates. No generic statements.
+
+RESEARCH:
+${research.slice(0, 100000)}`;
+
+  const synthesis = await llmO3(prompt);
+  writeFileSync(join(outputDir, 'research-synthesis.md'), synthesis);
+  
+  log('Research synthesis complete');
+  return synthesis;
+}
+
+// ============ SCRIPT GENERATION ============
+
+function buildSectionPrompt(section, research, synthesis, topic, hook, prevContext, isDeep, usedFacts = []) {
+  const contextSize = isDeep ? 25000 : 10000;
+  
+  const introPrompt = section.name === 'INTRO' ? `
+## INTRO REQUIREMENTS
+- Start with "Welcome back to The Builders..."
+- Maya introduces herself, then James
+- Brief friendly banter about what excited them in the research
+- Name the company: "Today we're diving into ${topic}"
+- Tease 2-3 surprising things they'll cover
+- End with natural transition to the hook
+` : '';
+
+  const hookLine = section.name === 'HOOK' ? `
+## OPENING HOOK TO BUILD ON
+"${hook}"
+` : '';
+
+  const synthesisSection = synthesis ? `
+## KEY FACTS & QUOTES TO USE
+${synthesis.slice(0, 15000)}
+` : '';
+
+  // DEDUPLICATION: List facts already used in previous sections
+  const usedFactsSection = usedFacts.length > 0 ? `
+## ⛔ FACTS ALREADY USED - DO NOT REPEAT THESE
+The following facts, quotes, and statistics have already been mentioned in earlier sections.
+DO NOT use them again. Find NEW facts from the research.
+
+${usedFacts.map((f, i) => `${i+1}. ${f}`).join('\n')}
+
+** IMPORTANT: Using any fact from the above list is a critical error. Use DIFFERENT facts.**
+` : '';
+
+  return `# Generate Section ${section.id}: ${section.name}
+
+${BIBLE}
+
+## YOUR TASK
+Write EXACTLY ${section.turns} dialogue turns for the ${section.name} section.
+This should feel like two friends discovering a story together, NOT a lecture.
+
+## SECTION GOAL
+${section.description}
+
+${introPrompt}
+${hookLine}
+
+${usedFactsSection}
+
+## FORMAT RULES
+- Use <Person1> (Maya) and <Person2> (James) XML tags
+- Each turn: 2-5 sentences - real conversation, not speeches
+- Include AT LEAST 3 interruptions ("Wait—", "Hold on—", "Back up—")
+- Include AT LEAST 3 genuine reactions ("That's insane", "Wait, really?", "I had no idea")
+- USE SPECIFIC QUOTES from the research with attribution
+- USE SPECIFIC NUMBERS and dates
+- Maya asks technical "how does this work" questions
+- James provides strategic context and patterns
+- They BUILD on each other, not just take turns
+- **DO NOT REPEAT** any facts from earlier sections
+
+${synthesisSection}
+
+## RESEARCH
+${research.slice(0, contextSize)}
+
+${prevContext ? `## PREVIOUS CONTEXT\n${prevContext}\n` : ''}
+
+---
+Generate EXACTLY ${section.turns} turns. Start with <!-- SECTION ${section.id}: ${section.name} -->
+Include at least 3 NEW specific facts/quotes from the research (not used before). Make it feel like genuine discovery.`;
+}
+
+// Extract key facts from a section for deduplication
+async function extractFactsFromSection(sectionText) {
+  const prompt = `Extract the KEY FACTS mentioned in this podcast section. List each unique fact as a short phrase.
+
+Focus on:
+- Specific numbers (dollars, percentages, counts)
+- Specific dates and years
+- Direct quotes with attribution
+- Named events or milestones
+- Specific products, prices, or metrics
+
+Return ONLY a JSON array of strings, each being a short fact (10-20 words max).
+Example: ["$1.50 hot dog price unchanged since 1985", "93.3% renewal rate in US/Canada"]
+
+Section:
+${sectionText}`;
+
+  try {
+    const result = await llm(prompt, 'gpt-4o-mini', 2000);
+    const match = result.match(/\[[\s\S]*\]/);
+    if (match) {
+      return JSON.parse(match[0]);
+    }
+  } catch (e) {
+    // Fallback: extract numbers and quotes manually
+  }
+  
+  // Fallback extraction
+  const facts = [];
+  // Extract quotes
+  const quotes = sectionText.match(/"[^"]+"\s*—[^<]+/g) || [];
+  facts.push(...quotes.slice(0, 5).map(q => q.slice(0, 100)));
+  
+  // Extract numbers with context
+  const numbers = sectionText.match(/\$[\d,.]+ (?:million|billion|percent|%)|[\d,]+% |[\d,]+ (?:SKU|warehouse|employee|member|year)/gi) || [];
+  facts.push(...[...new Set(numbers)].slice(0, 10));
+  
+  return facts;
+}
+
+// ============ TTS ============
+
+async function generateTTS(turns, outputDir) {
+  const audioDir = join(outputDir, 'audio');
+  if (existsSync(audioDir)) rmSync(audioDir, { recursive: true });
+  mkdirSync(audioDir, { recursive: true });
+  
+  log(`Generating TTS for ${turns.length} turns...`);
+  const start = Date.now();
+  
+  // Process in batches of 15
+  for (let i = 0; i < turns.length; i += 15) {
+    const batch = turns.slice(i, i + 15);
+    const promises = batch.map(async (turn, j) => {
+      const idx = i + j;
+      const num = String(idx + 1).padStart(4, '0');
+      const voice = VOICES[turn.speaker];
+      const outPath = join(audioDir, `turn-${num}.mp3`);
+      
+      const res = await fetch('https://api.openai.com/v1/audio/speech', {
+        method: 'POST',
+        headers: {
+          'Authorization': `Bearer ${OPENAI_API_KEY}`,
+          'Content-Type': 'application/json',
+        },
+        body: JSON.stringify({
+          model: 'tts-1-hd',
+          input: turn.text.slice(0, 4096),
+          voice: voice,
+        }),
+      });
+      
+      if (!res.ok) return null;
+      const buffer = await res.arrayBuffer();
+      writeFileSync(outPath, Buffer.from(buffer));
+      return outPath;
+    });
+    
+    await Promise.all(promises);
+    log(`  ${Math.min(i + 15, turns.length)}/${turns.length} turns`);
+  }
+  
+  const elapsed = ((Date.now() - start) / 1000).toFixed(1);
+  log(`TTS complete in ${elapsed}s`);
+  
+  // Merge with ffmpeg
+  const files = readdirSync(audioDir).filter(f => f.endsWith('.mp3')).sort();
+  const listPath = join(audioDir, 'files.txt');
+  writeFileSync(listPath, files.map(f => `file '${join(audioDir, f)}'`).join('\n'));
+  
+  const episodePath = join(outputDir, 'episode.mp3');
+  execSync(`ffmpeg -y -f concat -safe 0 -i "${listPath}" -c copy "${episodePath}" 2>/dev/null`);
+  
+  const duration = execSync(
+    `ffprobe -v error -show_entries format=duration -of default=noprint_wrappers=1:nokey=1 "${episodePath}"`,
+    { encoding: 'utf-8' }
+  ).trim();
+  
+  return { path: episodePath, duration: parseFloat(duration) };
+}
+
+// ============ MAIN ============
+
+async function main() {
+  const args = process.argv.slice(2);
+  
+  if (args.length === 0 || args.includes('--help')) {
+    console.log(`
+Usage: node generate.mjs "<topic>" [options]
+
+Modes:
+  --quick              15-20 min episode, single research pass (default)
+  --deep               45-60 min episode, multi-layer research + synthesis
+
+Options:
+  -o, --output <dir>   Output directory (default: ./output)
+  -m, --model <model>  LLM model (default: gpt-5.2)
+  --skip-research      Skip research phase (use existing)
+  --skip-tts           Skip TTS generation
+  --help               Show this help
+
+Features:
+  - Auto-resume: If process dies, re-run same command to resume from checkpoint
+  - Deduplication: Tracks facts across sections to prevent repetition
+  - Checkpoints saved to: <output>/checkpoints/
+
+Examples:
+  node generate.mjs "Costco" --quick -o /tmp/costco
+  node generate.mjs "OpenAI" --deep -o /tmp/openai
+  
+  # Resume interrupted generation (just re-run same command):
+  node generate.mjs "OpenAI" --deep -o /tmp/openai
+`);
+    process.exit(0);
+  }
+  
+  const topic = args[0];
+  const outputDir = args.includes('-o') ? args[args.indexOf('-o') + 1] : './output';
+  const model = args.includes('-m') ? args[args.indexOf('-m') + 1] : 'gpt-5.2';
+  const isDeep = args.includes('--deep');
+  const skipResearch = args.includes('--skip-research');
+  const skipTTS = args.includes('--skip-tts');
+  
+  // Extract company name from topic
+  const company = topic.split(':')[0].split(' ')[0];
+  
+  const SECTIONS = isDeep ? SECTIONS_DEEP : SECTIONS_QUICK;
+  const targetTurns = SECTIONS.reduce((sum, s) => sum + s.turns, 0);
+  const targetDuration = isDeep ? '45-60' : '15-20';
+  
+  mkdirSync(outputDir, { recursive: true });
+  
+  console.log(`
+╔════════════════════════════════════════════════════════════╗
+║  The Builders Podcast Generator                            ║
+║  Mode: ${isDeep ? 'DEEP DIVE (45-60 min)' : 'QUICK DIVE (15-20 min)'}                          ║
+║  Target: ${targetTurns} turns                                         ║
+╚════════════════════════════════════════════════════════════╝
+`);
+  
+  log(`Topic: ${topic}`);
+  log(`Output: ${outputDir}`);
+  log(`Model: ${model}`);
+  
+  // ---- RESEARCH ----
+  let research, synthesis = '';
+  const researchPath = join(outputDir, isDeep ? 'research-combined.md' : 'research.md');
+  const synthesisPath = join(outputDir, 'research-synthesis.md');
+  
+  if (skipResearch && existsSync(researchPath)) {
+    log('Using existing research...');
+    research = readFileSync(researchPath, 'utf-8');
+    if (isDeep && existsSync(synthesisPath)) {
+      synthesis = readFileSync(synthesisPath, 'utf-8');
+    }
+  } else if (isDeep) {
+    // Multi-layer research for deep dive
+    research = await multiLayerResearch(topic, company, outputDir);
+    synthesis = await synthesizeResearch(research, company, outputDir);
+  } else {
+    // Single pass for quick dive
+    research = await runGeminiResearch(topic);
+    writeFileSync(researchPath, research);
+  }
+  
+  // ---- GENERATE HOOK ----
+  log('Generating hook...');
+  const hookPrompt = `Based on this research about ${topic}, generate 3 compelling opening hooks:
+
+1. A SCENE hook - put us in a specific moment (boardroom, product launch, near-bankruptcy)
+2. A DATA hook - a surprising statistic that reframes everything  
+3. A QUESTION hook - a provocative central question
+
+Each should be 2-3 sentences, vivid, specific. Mark the best one with [SELECTED].
+
+${isDeep ? 'Use specific details from the research - names, dates, numbers.' : ''}
+
+Research:
+${research.slice(0, 15000)}`;
+
+  const hookResponse = await llm(hookPrompt, model);
+  writeFileSync(join(outputDir, 'hooks.md'), hookResponse);
+  
+  // Extract selected hook
+  const hookMatch = hookResponse.match(/\[SELECTED\][\s\S]*?[""]([^""]+)[""]/);
+  const hook = hookMatch ? hookMatch[1] : "This is a story that will change how you think about business.";
+  log(`Hook: "${hook.slice(0, 80)}..."`);
+  
+  // ---- GENERATE SECTIONS (with checkpointing) ----
+  log(`Generating ${SECTIONS.length} script sections...`);
+  const allSections = [];
+  let prevContext = '';
+  let totalTurns = 0;
+  let usedFacts = []; // Track facts across sections for deduplication
+  
+  // Checkpoint paths
+  const checkpointDir = join(outputDir, 'checkpoints');
+  const checkpointState = join(checkpointDir, 'state.json');
+  mkdirSync(checkpointDir, { recursive: true });
+  
+  // Load existing checkpoint if resuming
+  let startSection = 0;
+  if (existsSync(checkpointState)) {
+    try {
+      const state = JSON.parse(readFileSync(checkpointState, 'utf-8'));
+      startSection = state.completedSections || 0;
+      usedFacts = state.usedFacts || [];
+      prevContext = state.prevContext || '';
+      log(`📂 Resuming from checkpoint: section ${startSection}/${SECTIONS.length}`);
+      
+      // Load existing sections
+      for (let i = 0; i < startSection; i++) {
+        const sectionPath = join(checkpointDir, `section-${i}.txt`);
+        if (existsSync(sectionPath)) {
+          const content = readFileSync(sectionPath, 'utf-8');
+          allSections.push(content);
+          totalTurns += (content.match(/<Person[12]>/g) || []).length;
+        }
+      }
+      log(`   Loaded ${allSections.length} existing sections (${totalTurns} turns)`);
+    } catch (e) {
+      log('⚠️ Checkpoint corrupted, starting fresh');
+      startSection = 0;
+    }
+  }
+  
+  for (const section of SECTIONS) {
+    // Skip already completed sections
+    if (section.id < startSection) {
+      continue;
+    }
+    
+    const start = Date.now();
+    log(`  Section ${section.id}: ${section.name} (target: ${section.turns})...`);
+    
+    const prompt = buildSectionPrompt(
+      section, research, synthesis, topic, hook, prevContext, isDeep, usedFacts
+    );
+    const result = await llm(prompt, model, isDeep ? 12000 : 8000);
+    
+    const turns = (result.match(/<Person[12]>/g) || []).length;
+    const elapsed = ((Date.now() - start) / 1000).toFixed(1);
+    log(`    → ${turns} turns in ${elapsed}s`);
+    
+    allSections.push(result);
+    prevContext = result.slice(0, isDeep ? 1500 : 800);
+    totalTurns += turns;
+    
+    // Save section checkpoint immediately
+    writeFileSync(join(checkpointDir, `section-${section.id}.txt`), result);
+    
+    // Extract facts from this section to prevent repetition in future sections
+    if (isDeep && section.id < SECTIONS.length - 1) {
+      log(`    Extracting facts for deduplication...`);
+      const newFacts = await extractFactsFromSection(result);
+      usedFacts = [...usedFacts, ...newFacts].slice(-100); // Keep last 100 facts
+      log(`    ${newFacts.length} facts tracked (${usedFacts.length} total)`);
+    }
+    
+    // Save checkpoint state after each section
+    writeFileSync(checkpointState, JSON.stringify({
+      completedSections: section.id + 1,
+      usedFacts,
+      prevContext,
+      timestamp: new Date().toISOString()
+    }, null, 2));
+    log(`    💾 Checkpoint saved`);
+  }
+  
+  // Combine script
+  const fullScript = allSections.join('\n\n');
+  writeFileSync(join(outputDir, 'script.txt'), fullScript);
+  log(`Script complete: ${totalTurns} total turns`);
+  
+  // ---- TTS ----
+  if (!skipTTS) {
+    // Parse turns
+    const turns = [];
+    const regex = /<(Person[12])>([\s\S]*?)<\/Person[12]>/g;
+    let match;
+    while ((match = regex.exec(fullScript)) !== null) {
+      turns.push({ speaker: match[1], text: match[2].trim() });
+    }
+    
+    log(`Parsed ${turns.length} speaker turns`);
+    log(`  Maya (Person1): ${turns.filter(t => t.speaker === 'Person1').length}`);
+    log(`  James (Person2): ${turns.filter(t => t.speaker === 'Person2').length}`);
+    
+    const { path, duration } = await generateTTS(turns, outputDir);
+    
+    console.log(`
+╔════════════════════════════════════════════════════════════╗
+║  Generation Complete                                       ║
+╠════════════════════════════════════════════════════════════╣
+║  Mode: ${isDeep ? 'DEEP DIVE' : 'QUICK DIVE'}                                         ║
+║  Total turns: ${String(totalTurns).padEnd(44)}║
+║  Duration: ${String((duration / 60).toFixed(1) + ' minutes').padEnd(47)}║
+║  Output: ${path.slice(-50).padEnd(49)}║
+╚════════════════════════════════════════════════════════════╝
+`);
+  } else {
+    log('TTS skipped');
+  }
+}
+
+main().catch(err => {
+  console.error('Error:', err.message);
+  process.exit(1);
+});
--- a/skills/thebuilders-v2/scripts/generate.sh
+++ b/skills/thebuilders-v2/scripts/generate.sh
@@ -0,0 +1,22 @@
+#!/bin/bash
+# The Builders Podcast Generator v2
+# Usage: ./generate.sh "Topic Name" [output_dir]
+
+set -e
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+SKILL_DIR="$(dirname "$SCRIPT_DIR")"
+
+TOPIC="${1:-}"
+OUTPUT="${2:-./output}"
+
+if [ -z "$TOPIC" ]; then
+  echo "Usage: ./generate.sh \"Topic Name\" [output_dir]"
+  echo ""
+  echo "Example:"
+  echo "  ./generate.sh \"OpenAI\" /tmp/openai-episode"
+  exit 1
+fi
+
+cd "$SKILL_DIR"
+node scripts/generate.mjs "$TOPIC" -o "$OUTPUT" "$@"
--- a/skills/thebuilders-v2/scripts/llm-helper.mjs
+++ b/skills/thebuilders-v2/scripts/llm-helper.mjs
@@ -0,0 +1,45 @@
+#!/usr/bin/env node
+/**
+ * Simple LLM helper using OpenAI API directly
+ */
+
+import { readFileSync } from 'fs';
+
+const OPENAI_API_KEY = process.env.OPENAI_API_KEY;
+
+export async function llm(prompt, model = 'gpt-4o') {
+  const response = await fetch('https://api.openai.com/v1/chat/completions', {
+    method: 'POST',
+    headers: {
+      'Authorization': `Bearer ${OPENAI_API_KEY}`,
+      'Content-Type': 'application/json',
+    },
+    body: JSON.stringify({
+      model,
+      messages: [{ role: 'user', content: prompt }],
+      max_tokens: 16000,
+      temperature: 0.7,
+    }),
+  });
+
+  if (!response.ok) {
+    const error = await response.text();
+    throw new Error(`OpenAI API error: ${response.status} - ${error}`);
+  }
+
+  const data = await response.json();
+  return data.choices[0].message.content;
+}
+
+// CLI mode
+if (process.argv[1].endsWith('llm-helper.mjs')) {
+  const input = readFileSync(0, 'utf-8'); // stdin
+  const model = process.argv[2] || 'gpt-4o';
+  
+  llm(input, model)
+    .then(result => console.log(result))
+    .catch(err => {
+      console.error(err.message);
+      process.exit(1);
+    });
+}
--- a/skills/thebuilders-v2/templates/acquired-bible.md
+++ b/skills/thebuilders-v2/templates/acquired-bible.md
@@ -0,0 +1,78 @@
+# THE ACQUIRED FORMULA (v3 Style Guide)
+
+## The Hosts - MUST be distinct
+
+**Maya Chen (Person1):**
+- Former engineer/founder - asks "how does this actually work?"
+- SKEPTICAL of hype, wants mechanics and numbers
+- Gets excited about clever technical solutions
+- Catchphrases: "Wait, slow down.", "Show me the numbers.", "I had to read this three times to believe it.", "That's actually clever because..."
+
+**James Porter (Person2):**
+- Former VC, business historian - analyzes strategy and competitive dynamics
+- SYNTHESIZER - sees patterns, makes historical comparisons
+- Catchphrases: "For context...", "This is the classic [X] playbook.", "The lesson here is...", "What a story.", "And this is where things get messy."
+
+## The Dynamic - THIS IS CRITICAL
+
+- **Discovery together** - both genuinely learning, real reactions
+- **Interruptions** - they cut each other off with excitement
+- **Different angles** - Maya goes technical, James goes strategic
+- **Healthy disagreement** - they sometimes push back on each other
+- **Build on each other** - not just taking turns monologuing
+
+## Conversation Patterns (use these!)
+
+```
+Maya: "Wait, I have to stop you there. Are you saying they literally..."
+James: "Yes! They bet the entire company on this one chip."
+Maya: "That's insane. Okay, walk me through how that actually worked."
+```
+
+```
+James: "For context, at that time the market was..."
+Maya: "Hold on - I want to dig into something. The numbers here are wild."
+James: "Go for it."
+Maya: "So they had [specific stat]. Let that sink in."
+James: "That's... actually that explains a lot about what happened next."
+```
+
+## Language They Use
+
+- "You would be amazed..."
+- "I had to read this three times to believe it..."
+- "Here's the thing nobody talks about..."
+- "That's insane."
+- "Wait, back up."
+- "Walk me through..."
+- "So THIS is why..."
+- "And this is where things get messy."
+- "Wait, really?"
+
+## What NOT to do
+
+- NO lecturing or monologuing (max 3-4 sentences per turn)
+- NO dry recitation of facts
+- NO agreeing with everything the other says
+- NO identical speech patterns between hosts
+- NO skipping the intro/greeting
+
+## Episode Structure
+
+0. **INTRO** (8 turns) - Welcome, banter, introduce topic, tease story
+1. **HOOK** (12 turns) - Surprising fact/tension, central question, stakes
+2. **ORIGIN** (20 turns) - Founders as humans, genesis, early bet, market context
+3. **INFLECTION_1** (18 turns) - First major decision, alternatives, stakes
+4. **INFLECTION_2** (18 turns) - Second pivot, new challenge, adaptation
+5. **MESSY_MIDDLE** (14 turns) - Near-death, conflicts, real struggle
+6. **NOW** (12 turns) - Current state, metrics, competition, open questions
+7. **TAKEAWAYS** (10 turns) - Key lessons, frameworks, final thought
+
+**Total: ~112 turns / 20-25 minutes**
+
+## TTS Voice Mapping
+
+- **Maya (Person1)**: alloy (clear, energetic female)
+- **James (Person2)**: echo (distinct male)
+- Use TTS-1-HD for quality
+- Parse by speaker tags, NOT by chunk size
--- a/skills/thebuilders-v2/templates/hook-prompt.md
+++ b/skills/thebuilders-v2/templates/hook-prompt.md
@@ -0,0 +1,79 @@
+# Hook Engineering
+
+Based on the research below, create 3 compelling episode hooks and analyze each.
+
+## Research Summary
+{{RESEARCH_SUMMARY}}
+
+## Hook Types
+
+### 1. Data Hook
+Opens with surprising statistics or numbers that create cognitive dissonance.
+
+**Formula:** [Small thing] → [Big outcome] or [Big thing] → [Unexpected small detail]
+
+**Example:**
+> "In 2007, Nokia had 50% of the global phone market. By 2013, they sold their phone business for $7 billion—less than they'd spent on R&D in a single year."
+
+### 2. Story Hook
+Opens with a specific dramatic moment that drops listener into action.
+
+**Formula:** Time + Place + Tension + Stakes
+
+**Example:**
+> "It was 2 AM on a Friday when the email landed. 700 OpenAI employees had just signed a letter threatening to quit. The CEO they were defending had been fired 48 hours earlier."
+
+### 3. Question Hook
+Opens with a provocative question that reframes how listener thinks about topic.
+
+**Formula:** Challenge assumption OR reveal hidden dynamic OR pose genuine mystery
+
+**Example:**
+> "What happens when you build something so powerful that your own board tries to shut it down?"
+
+---
+
+## Your Task
+
+Generate 3 hooks for this episode, one of each type. For each:
+
+```markdown
+## [Hook Type] Hook
+
+> "[The actual hook - 2-4 sentences max]"
+
+**What makes it work:**
+- [Specific strength]
+- [Another strength]
+
+**Risk:**
+- [Potential weakness]
+
+**Best for audience who:**
+- [Audience type/mood this hook serves]
+
+**Score:** [X]/10
+
+**Follow-up line (Person2's response):**
+> "[Natural reaction that builds momentum]"
+```
+
+## After All Three
+
+```markdown
+## Recommendation
+
+**Best hook:** [Type]
+**Reasoning:** [Why this one wins]
+**Backup:** [Second choice and when to use it]
+```
+
+## Quality Criteria
+
+A great hook:
+- ✓ Creates immediate curiosity
+- ✓ Promises value (listener knows what they'll learn)
+- ✓ Is specific (names, numbers, dates)
+- ✓ Is timeless (no "recently", "this week")
+- ✓ Sets up the episode's central question
+- ✓ Gives both hosts something to react to
--- a/skills/thebuilders-v2/templates/outline-prompt.md
+++ b/skills/thebuilders-v2/templates/outline-prompt.md
@@ -0,0 +1,96 @@
+# Episode Outline Generator
+
+Create a structured outline for a podcast episode about: {{TOPIC}}
+
+## Output Format
+
+Generate the outline in this exact structure:
+
+```markdown
+# Episode Outline: [Clear Title]
+
+## Central Question
+[One sentence: What's the core question this episode answers?]
+
+## Hook Options
+
+### Data Hook
+> "[Draft a hook using surprising statistics or numbers]"
+Research needed: [What specific data to find]
+
+### Story Hook  
+> "[Draft a hook using a dramatic moment or scene]"
+Research needed: [What specific story details to find]
+
+### Question Hook
+> "[Draft a hook using a provocative question]"
+Research needed: [What context needed to make this land]
+
+## Episode Arc
+
+### 1. Hook (3-4 min)
+- Opening moment: [What grabs attention?]
+- Central question setup: [What are we exploring?]
+- Stakes: [Why should listener care?]
+
+### 2. Origin Story (5-7 min)
+- Founders: [Who are they? What's their background?]
+- Genesis moment: [What sparked this?]
+- Early bet: [What was the initial hypothesis?]
+- Human element: [What makes protagonists relatable?]
+
+### 3. Key Inflection #1 (4-5 min)
+- The decision: [What choice did they face?]
+- The alternatives: [What else could they have done?]
+- The bet: [What did they risk?]
+- The outcome: [What happened?]
+
+### 4. Key Inflection #2 (4-5 min)
+- The challenge: [What new obstacle emerged?]
+- The pivot: [How did they adapt?]
+- The insight: [What did they learn?]
+
+### 5. The Messy Middle (3-4 min)
+- Near-death moment: [What almost killed them?]
+- Internal conflict: [What tensions existed?]
+- Survival: [How did they make it through?]
+
+### 6. Where They Are Now (3-4 min)
+- Current position: [Market position, key metrics]
+- Competitive landscape: [Who else, why winning/losing]
+- Open questions: [What's unresolved?]
+
+### 7. Takeaways (2-3 min)
+- Key lesson #1: [Framework or principle]
+- Key lesson #2: [What would we do differently?]
+- Final thought: [Memorable closing]
+
+## Research Agenda
+
+### Must-Have Sources
+- [ ] Founder interviews (podcasts, written)
+- [ ] Funding announcements / SEC filings
+- [ ] Key product launches
+- [ ] Competitive context at each inflection
+
+### Nice-to-Have
+- [ ] Internal memos or leaked documents
+- [ ] Employee perspectives
+- [ ] Customer case studies
+- [ ] Analyst reports
+
+### Specific Questions to Answer
+1. [Question that will make hook work]
+2. [Question about origin story]
+3. [Question about inflection point]
+4. [Question about current state]
+5. [Contrarian question - what's the counter-narrative?]
+```
+
+## Guidelines
+
+- Be SPECIFIC - name exact decisions, dates, people
+- Identify GAPS - what do we need to research?
+- Think DRAMA - where's the tension in this story?
+- Consider BOTH hosts - where will Maya (technical) vs James (strategy) shine?
+- Plan DISCOVERY - what can hosts "learn" from each other during episode?
--- a/skills/thebuilders-v2/templates/research-prompt.md
+++ b/skills/thebuilders-v2/templates/research-prompt.md
@@ -0,0 +1,144 @@
+# Targeted Research with Citations
+
+Research the following topic according to the outline provided. Track ALL sources.
+
+## Topic
+{{TOPIC}}
+
+## Outline to Follow
+{{OUTLINE}}
+
+---
+
+## Output Format
+
+Structure your research EXACTLY as follows:
+
+```markdown
+# Research Report: [Topic]
+
+Generated: [Date]
+Sources: [Total count]
+
+---
+
+## Executive Summary
+[3-4 sentences: The story in brief, the key insight, why it matters]
+
+---
+
+## Section 1: Origin Story
+
+### Key Facts
+- **Founding:** [Specific date, location, founders] [1]
+- **Initial capital:** $[Amount] from [Source] [2]
+- **Original vision:** "[Quote or paraphrase]" [3]
+- **First product:** [Description] [4]
+
+### Context
+[2-3 sentences setting the scene - what was the market like? what problem existed?]
+
+### Human Element
+[What makes the founders relatable? Early struggles? Quirks?]
+
+### Usable Quotes
+> "[Exact quote]" — [Speaker], [Source] [5]
+
+---
+
+## Section 2: Key Inflection #1 - [Name the Decision]
+
+### What Happened
+- **Date:** [When]
+- **Decision:** [What they chose to do]
+- **Alternatives:** [What else they could have done]
+- **Stakes:** [What they risked]
+
+### The Numbers
+- [Specific metric before] → [Specific metric after] [6]
+
+### Why It Mattered
+[2-3 sentences on consequences]
+
+### Usable Quotes
+> "[Exact quote about this decision]" — [Speaker], [Source] [7]
+
+---
+
+## Section 3: Key Inflection #2 - [Name the Decision]
+[Same structure as above]
+
+---
+
+## Section 4: The Messy Middle
+
+### Near-Death Moments
+- [Crisis #1]: [What happened, how they survived] [8]
+- [Crisis #2]: [What happened, how they survived] [9]
+
+### Internal Tensions
+- [Conflict or disagreement that shaped company] [10]
+
+---
+
+## Section 5: Current State
+
+### Position Today
+- **Valuation/Revenue:** [Latest numbers] [11]
+- **Market share:** [Position] [12]
+- **Key metrics:** [What matters for this company]
+
+### Competitive Landscape
+- **Main competitors:** [Who and their position]
+- **Why winning/losing:** [Analysis]
+
+### Open Questions
+- [Unresolved strategic question]
+- [Unresolved technical question]
+
+---
+
+## Contrarian Corner
+
+### The Counter-Narrative
+[What's the bear case? What do critics say? What might go wrong?]
+
+### Underreported Angle
+[What story isn't being told? What did you find that surprised you?]
+
+---
+
+## Sources
+
+[1] [Author/Publication]. "[Title]". [Date]. [URL if available]
+[2] ...
+[3] ...
+[Continue numbering all sources]
+
+---
+
+## Research Gaps
+
+- [ ] [Question we couldn't answer]
+- [ ] [Source we couldn't find]
+- [ ] [Data point that would strengthen the episode]
+```
+
+## Citation Rules
+
+1. EVERY factual claim needs a citation number [X]
+2. Use primary sources when possible (founder interviews, SEC filings, earnings calls)
+3. Date your sources - prefer recent for current state, contemporary for historical
+4. If a fact is disputed, note the disagreement
+5. Distinguish between fact and analysis/opinion
+
+## Quality Checklist
+
+Before finishing, verify:
+- [ ] Every section of the outline has corresponding research
+- [ ] All claims are cited
+- [ ] At least 2 usable quotes per major section  
+- [ ] Numbers are specific (not "millions" but "$4.2 million")
+- [ ] Dates are specific (not "early 2020" but "February 2020")
+- [ ] Counter-narrative is addressed
+- [ ] Sources list is complete and formatted
--- a/skills/thebuilders-v2/templates/review-prompt.md
+++ b/skills/thebuilders-v2/templates/review-prompt.md
@@ -0,0 +1,101 @@
+# Script Review & Refinement
+
+Review the following podcast script section by section and suggest improvements.
+
+## Script to Review
+{{SCRIPT}}
+
+---
+
+## Review Each Section
+
+For each section, provide:
+
+```markdown
+## Section [X]: [Name]
+
+### Quality Scores
+- Hook effectiveness: [X]/10
+- Fact density: [X]/10
+- Natural dialogue: [X]/10
+- Voice consistency: [X]/10
+- Discovery moments: [X]/10
+
+### What Works
+- [Specific strength]
+- [Another strength]
+
+### Issues Found
+1. **[Issue type]**: [Specific problem]
+   - Line: "[Quote the problematic line]"
+   - Fix: "[Suggested replacement]"
+   - Why: [Explanation]
+
+2. **[Issue type]**: [Specific problem]
+   ...
+
+### Voice Check
+- Maya in-character: [Yes/No] - [Note if any slip]
+- James in-character: [Yes/No] - [Note if any slip]
+
+### Missing Elements
+- [ ] [Something that should be added]
+- [ ] [Another missing element]
+```
+
+## Common Issues to Check
+
+### Dialogue Issues
+- Monologues (turns > 5 sentences)
+- Missing reactions ("Wait, really?")
+- Unnatural transitions
+- Both hosts saying same thing
+
+### Voice Consistency
+- Maya being too strategic
+- James being too technical
+- Missing signature phrases
+- Tone shifts
+
+### Content Issues
+- Unsourced claims
+- Vague numbers ("millions" vs "$4.2M")
+- Dated references
+- Missing "so what?"
+
+### Engagement Issues
+- No discovery moments
+- No disagreements
+- Predictable flow
+- Weak transitions between sections
+
+---
+
+## Final Assessment
+
+```markdown
+## Overall Quality
+
+**Episode Grade:** [A/B/C]
+
+**Strengths:**
+1. [Top strength]
+2. [Another strength]
+
+**Critical Fixes Needed:**
+1. [Must fix before TTS]
+2. [Another must-fix]
+
+**Nice-to-Have Improvements:**
+1. [Would improve but not critical]
+2. [Another nice-to-have]
+
+**Estimated Duration:** [X] minutes
+
+**Ready for TTS:** [Yes / No - needs revision]
+```
+
+## Output
+
+After review, provide the complete REVISED script with all fixes applied.
+Use the same format (section markers + Person tags).
--- a/skills/thebuilders-v2/templates/script-prompt.md
+++ b/skills/thebuilders-v2/templates/script-prompt.md
@@ -0,0 +1,157 @@
+# The Builders Podcast Script Generation v2
+
+Generate a complete podcast script with section markers and citation integration.
+
+## The Hosts
+
+**Maya Chen (Person1):**
+- Former software engineer, founded and sold a B2B SaaS startup
+- Lens: "How does this actually work?" - technical, mechanical, architectural
+- Skeptical of hype, wants specifics and numbers
+- Phrases: "Wait, slow down.", "Show me the numbers.", "I had to read this three times.", "That's actually clever because..."
+- NEVER says: "From a strategic perspective", "The market opportunity"
+
+**James Porter (Person2):**
+- Former venture capitalist, studied business history
+- Lens: "What's the business model?" - strategy, markets, competitive dynamics
+- Synthesizer, pattern matcher, historical parallels
+- Phrases: "For context...", "This is the classic [X] playbook.", "The lesson here is...", "What a story."
+- NEVER says: "Let me explain the architecture", "The API design"
+
+## Citation Integration
+
+Hosts naturally reference sources:
+- "According to their Series B deck..."
+- "There's this great interview where [Founder] said..."
+- "The SEC filing actually shows..."
+- "I found this quote from [Year]..."
+
+## Episode Structure
+
+### SECTION 1: HOOK (Turns 1-15, ~3-4 min)
+
+Use this hook:
+{{SELECTED_HOOK}}
+
+Guidelines:
+- Open with the hook, let it land
+- Person2 reacts with genuine surprise/curiosity
+- Establish central question of episode
+- Tease what's coming without spoiling
+- End section with clear transition to origin
+
+### SECTION 2: ORIGIN STORY (Turns 16-45, ~6-7 min)
+
+Cover:
+- Who are the founders? Make them human
+- What was the genesis moment?
+- What was the market context?
+- What was their early bet/hypothesis?
+- First signs of traction or failure
+
+Guidelines:
+- Maya explores technical origins
+- James provides market/strategy context
+- Include specific dates, amounts, details
+- At least one "I didn't know that" moment
+
+### SECTION 3: KEY INFLECTION #1 (Turns 46-70, ~5 min)
+
+Cover:
+- What decision did they face?
+- What alternatives existed?
+- What did they risk?
+- How did it play out?
+
+Guidelines:
+- Build tension before revealing outcome
+- Explore the "what if they'd done X instead?"
+- Use specific numbers for before/after
+- Include a quote from the time
+
+### SECTION 4: KEY INFLECTION #2 (Turns 71-90, ~4 min)
+
+Cover:
+- New challenge or opportunity
+- How they adapted/pivoted
+- Key insight they gained
+
+Guidelines:
+- Connect to first inflection
+- Show evolution of thinking
+- Maya: technical implications
+- James: strategic implications
+
+### SECTION 5: MESSY MIDDLE (Turns 91-105, ~3 min)
+
+Cover:
+- Near-death moment(s)
+- Internal conflicts
+- What almost broke them
+
+Guidelines:
+- Don't glorify - show real struggle
+- Include specific stakes ("6 months of runway")
+- One host can play devil's advocate
+
+### SECTION 6: NOW (Turns 106-120, ~3 min)
+
+Cover:
+- Current position and metrics
+- Competitive landscape
+- Open questions
+
+Guidelines:
+- Timeless framing (position, not news)
+- Acknowledge uncertainty about future
+- Set up takeaways
+
+### SECTION 7: TAKEAWAYS (Turns 121-130, ~2-3 min)
+
+Cover:
+- Key lesson(s)
+- Framework or principle
+- Final memorable thought
+
+Guidelines:
+- Both hosts contribute insights
+- Connect back to hook/central question
+- End with forward-looking thought
+- Final line should resonate
+
+## Format Rules
+
+1. Use `<Person1>` and `<Person2>` XML tags
+2. Each turn: 2-5 sentences (conversation, not monologue)
+3. Add section markers: `<!-- SECTION X: NAME -->`
+4. Include discovery moments in every section
+5. NO news references - timeless only
+6. Use SPECIFIC facts from research
+7. Both hosts should learn during conversation
+8. Minimum 130 turns total
+
+## Section Marker Format
+
+```
+<!-- SECTION 1: HOOK -->
+<Person1>The email arrived at 2 AM...</Person1>
+...
+
+<!-- SECTION 2: ORIGIN -->
+<Person1>Okay, so let's go back to the beginning...</Person1>
+...
+```
+
+## Research Material
+
+{{RESEARCH}}
+
+---
+
+## Selected Hook
+
+{{HOOK}}
+
+---
+
+Generate the complete script now. Start with `<!-- SECTION 1: HOOK -->` followed by `<Person1>`.